Do university AI policies actually protect what credentials mean?

The paper audits "verified public generative AI assessment guidance from 30 universities" and reports that policies are "becoming better at classifying AI use than at explaining what evidence and protections preserve credential validity." Boundaries are more visible than evidence standards, safeguards are uneven, and guidance is clearest when AI use resembles final-output substitution rather than feedback, access, verification, or professional workflow. The universities were not silent: many had official pages, AI-use categories, and disclosure language. The gap the authors identify is specific, since "rules about allowed use outpaced evidence for what credentials still certify." Hence the conclusion that permission categories are "necessary but insufficient."

The reasoning starts from a premise that generative AI has made "newly fragile": that a submitted artifact can stand as evidence of a learner's competence. The paper argues that the question is "not whether AI touched the work, but which cognitive operations moved from the learner to the system and which remained with the learner." Its example contrasts a student who uses AI feedback while keeping problem formulation, source evaluation, revision judgment, and final responsibility with one who delegates topic selection, evidence search, argument structure, drafting, citation, and prose revision. Its proposed framework, cognitive stewardship, links four things before a product is treated as evidence of competence: the learning claim, the delegation boundary, the evidence standard, and the safeguard layer. A permission category names the boundary. My reading is that it leaves the other three unstated: which retained operations the credential claims, what would show the learner still performed them, and what protects that evidence.

This sits close to several notes about the distance between an AI-assisted artifact and the human capacity behind it. Does AI assistance help workers learn lasting skills? supplies the reason an artifact is weak evidence: performance with assistance can be higher than what the person can do alone. This paper moves that problem from the individual worker to the certifying institution. Does AI assistance actually harm the way developers learn? distinguishes delegation patterns by how much cognitive engagement they keep, which resembles the paper's delegation boundary, though the excerpt does not cite or test that link. Can AI verify research outputs as fast as it generates them? describes verification lagging generation, and the evidence standard is the credential's version of verification. Disclosure language also touches Do users truly own the AI-generated content they produce?: if felt and claimed authorship diverge, a declaration of AI use may not establish who did the retained work.

The excerpt gives no actual scores, no selection criteria for the 30 universities beyond "verified public" guidance, and no agreement figures for the four open-weight LLM coders or any check of them against human coders. It says the scores were averaged "to reduce dependence on any single model's bias," but reports nothing about how well that worked. It does not test whether cognitive stewardship improves the validity of any credential, and it audits what institutions publish, not what they do. The finding is therefore best read as a documentation gap in published policy, and the framework as a proposal.